Impact of Synaptic Device Variations on Classification Accuracy in a Binarized Neural Network

Area of Science:

• Neuromorphic engineering
• Artificial intelligence hardware

Background:

• Neuromorphic systems require reliable synaptic devices for cognitive tasks.
• Analog synaptic devices face sustainability issues due to variability.
• Digital-type switching devices offer a more reliable alternative.

Purpose of the Study:

• To quantitatively investigate the impact of device parameter variations on BNN classification accuracy.
• To demonstrate the feasibility of neuromorphic systems using mature digital technologies.

Main Methods:

• Simulated BNNs for facial image classification using a supervised training scheme.
• Analyzed effects of weight states (Nstate), update margin (ΔG), and update variation (Gvar) on accuracy.

Main Results:

• Device parameter variations were quantitatively assessed for their effect on classification accuracy.
• The study confirmed the feasibility of BNNs for practical neuromorphic applications.

Conclusions:

• BNNs utilizing mature digital devices are a viable approach for practical neuromorphic systems.
• This research validates the use of digital-type switching devices in energy-efficient computing architectures.